Refit with less training data and longer horizons

[ethanwhite]

We want to be able to understand how the length of the forecast horizon plays into model performance. Since most of our models basically no large scale temporal structure we should be able to fit these models on shorter time periods allowing us to test against longer horizons. My initial thought is 10 years of training data, 25 years for forecast evaluation.

[ethanwhite]

Ideally #251 can be finished before this is run so that we have the maximum time-series length possible.